topological data analysis and stochastic topology - … · topological data analysis and stochastic...

topological data analysisand stochastic topology

yuliy baryshnikovwaikiki, march 2013

Promise of topological data analysis: extract the structure from the data.

In: point clouds

Out: hidden structure

Underlying structure: manifolds?stratified sets?

Two ways to look at them: Poincare's and Riemann:

Correspondingly, there are two approaches in dimensionality reduction:

• Poincare view - use ambient coordinates - PCA, Kernel PCA,...

• Riemann view - only intrinsic proximity makes sense - IsoMap, EigenMap,...

Topological Data Analysis aims at recovery of the most stable invariants of the data.

In particular, complete scale invariance is sought

This is both a curse and a blessing:

● very susceptible to noise● only vague notion of proximity matters

Invariants of topological spaces: two spaces are the same if they are

homeomorphic?

diffeomorphic?

homotopic?

Invariants of topological spaces:

Dimension?

Homotopy groups?

Homology and cohomology

Homology and cohomology

come in different flavors and colors

singular, de Rham, simplicial

Simplicial - easiest to deal with. Works well with meshes...

Algebraic formalism includes chains, cycles and boundaries, and the idiosyncratic relation

In essence, linear algebra encoding combinatorics of adjacencies.

Homology groups depend on the coefficient ring: what are the entries of the boundary matrix.

Sometimes there is torsion:

Working over a field – ranks of the homology groups are called Betti numbers: numeric invariants of a topological space

Morse theory – an important tool to generate the cellular partition of the underlying manifold, and estimate the Betti numbers

What happens when one adds a simplex of dimension d: either a cycle in dimension d-1 disappears, or a cycle of dimension d appears.

This implies existence of an aggregate invariant, the Euler characteristic.

Back to data: given a point cloud, how to arrive at a topological space? Taking into account desirable scale?

Old idea: alpha-shapes (Robbins, Edelsbrunner...)

Place a ball around each point in the dataset and construct the Čech complex: glue in a simplex for any non-empty intersection of the balls.

Nerve Theorem: If all intersections are contractible, the union of balls is homotopy equivalent to the Čech complex.

Noise is a problem, however: how to get rid of the small features?Answer: increase the scale.and againand againUntil the features disappear altogether

Persistence is a way to address the onslaught of irrelevant features

Tidal traces to be ignored: only the islands survive.

Persistence homology

Start with a filtration, and record when a cycle appears, and when it dies.

Encoding: either barcodes, or persistence diagram

Important result: stability theorem – small perturbations of the filtration change the persistence diagram little

Important tool: Morse theory

Some examples

Mumford dataset and the Klein bottle

Krioukov et al. Hyperbolic mappingof complex networks

Simple test: what is the dimension of the internet?

How we find the dimension knowing a sample from a manifold?

Testing on a special (stratified) set

Testing on a special (stratified) set

Autonomous Systems (“Internet”)

Stochastic Topology: dealing with the baseline case

Niyogi, Smale, Weinberger: Cech complex for small enough r and dense enough sample with has the homotopy type of the underlying manifold

Big Picture: from dust to foam to monolith

Big Picture: timescale

Big Picture: foam phase

Foam phase: oscillation of Euler characteristic

Euler characteristic in dim=7

Foam phase: concentrations of Betti numbers

Life in the ambient space: crackle phenomenon

If the noise is Gaussian, the balls form a solid ball. In exponential and subexponential cases, the homologies form in layers.

Plenty of further models and questions in stochastic topology

Persistence diagrams for Brownian bridge

Plenty of further models and questions in stochastic topology

Topology of the level sets of random functions

Random simplicial complexes

Random configuration spaces

...